Generative AI is rapidly entering classrooms and learning platforms, yet evidence about its impact on learning remains mixed. In this talk, I present a framework for making GenAI work in education through two complementary lenses - knowledge engineering (to increase cognitive fidelity in what AI evaluates and generates) and cognitively aligned interface design (to preserve learners' meaningful engagement). First, I present a randomized controlled trial with 354 students examining the impact of AI-mediated feedback on students' disciplinary writing performance and learning, compared to human-only feedback. We introduce and evaluate FeedbackWriter, a system that generates rubric-level AI suggestions to teaching assistants (TAs) as they provide feedback on students' economics essays. Students who received FeedbackWriter-supported feedback produced higher quality revisions than those who received TA-only feedback. This story illustrates how knowledge engineering can enhance cognitive fidelity and enable reliable feedback generation. Second, I illustrate cognitively-aligned interface (CAI) design with FeedbackWriter and a second system NoteCopilot. The FeedbackWriter interface aligns with TAs' cognitive processes for feedback provision, whereas NoteCopilot preserves cognitive engagement when learners use AI to take notes. In a controlled study comparing multiple levels of AI assistance in NoteCopilot, we find that an intermediate level of AI support can reduce extraneous cognitive load while preserving the necessary cognitive engagement to encode information. I use these two stories to discuss how GenAI may shift instructional labor and classroom resources, and how to navigate the AI assistance dilemma.